MOSS is a tool-augmented conversational language model developed by Fudan University that enables dialogue systems with external tool integration capabilities.
The project addresses the limitations of standard language models by augmenting them with access to external tools such as search engines, image generation, calculators, and equation solvers. The approach involves training a base model on high-quality bilingual text corpora, then fine-tuning it on multi-turn dialogue data to develop instruction-following and multi-turn conversation abilities. The tool-augmented variants are further trained on dialogue data enhanced with plugin interactions, allowing the model to learn when and how to invoke external tools during conversations.
Developers should consider MOSS if they need an open-source conversational model with explicit tool-use capabilities rather than relying solely on the model's internal knowledge. The project provides multiple model variants optimized for different hardware constraints, including quantized versions that require substantially less GPU memory than the full-precision models. This makes the tool accessible for deployment on more modest hardware setups. The project is particularly suited for applications requiring factual accuracy through search integration, dynamic computation, or image generation capabilities within a conversational interface.
The project maintains active development with regular model releases and refinements. The team has published research documenting the approach and continues to expand the model lineup with improved variants trained using preference feedback. The codebase includes comprehensive documentation for local deployment, fine-tuning procedures, and usage examples across different model configurations. Community engagement is facilitated through multiple language documentation and official communication channels.
